Conversation marketing strategy optimization method and system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent dialogues, and discloses a dialogue marketing strategy optimization method based on artificial intelligence, and the method comprises the following steps: collecting the voice, text, facial expression and physiological signals of a user in real time through a multi-modal perception assembly of a terminal device, constructing a dynamic emotion map, and extracting a multi-modal feature vector; a cross-modal information processing module is utilized to align the multi-modal data through a comparative learning algorithm, and causal relationship description of user behaviors and strategies and strategy risk scores are generated; cooperatively training a global strategy model through differential privacy and homomorphic encryption technologies; combining the dynamic emotion map and a causal model to generate an emotion adaptive dialogue script, and optimizing strategy selection through a reinforcement learning algorithm; and dynamically updating the emotion map, the risk score and the global model through a closed-loop feedback mechanism to form a real-time optimized strategy generation system. According to the invention, the practicability of artificial intelligence to real-time services can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent dialogue, and in particular to a method and system for optimizing marketing strategies based on artificial intelligence dialogue. Background Art

[0002] In recent years, the application of artificial intelligence (AI) in commercial marketing has evolved from single-function tools to complex integrated systems. Conversational AI leverages natural language processing to enable 24 / 7 customer service, significantly improving interaction efficiency. Predictive analytics and personalized recommendation systems optimize advertising and product recommendations based on user behavior data (such as click and purchase history), driving increased conversion rates.

[0003] However, existing technologies still have limitations in complex marketing scenarios: on the one hand, the fragmentation of multimodal data leads to insufficient dynamic modeling of users' emotional states, making it difficult to capture subconscious behavioral characteristics; on the other hand, there is a lack of coordination between privacy protection mechanisms (such as differential privacy) and strategy optimization goals, making it difficult to balance the efficiency and security of cross-enterprise data collaboration.

[0004] These problems limit the in-depth application of artificial intelligence in the marketing field, resulting in delayed model updates and difficulty in adapting to real-time business needs. As can be seen from the above, how to improve the practicality of artificial intelligence for real-time business still needs to be solved. Summary of the Invention

[0005] In order to improve the practicality of artificial intelligence for real-time business, this application provides a method and system for optimizing marketing strategies based on artificial intelligence dialogue.

[0006] In the first aspect, this application provides a method for optimizing marketing strategies based on artificial intelligence dialogue, which adopts the following technical solutions: A method for optimizing marketing strategies based on artificial intelligence dialogue, comprising: The multimodal perception component deployed on the terminal device collects multiple rounds of interaction data between the user and the AI ​​dialogue system in real time. The multi-round interaction data includes voice signals, text data, facial expression signals, and physiological signals. Based on the multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series, and the multimodal feature vectors corresponding to the multi-round interaction data are extracted. The multimodal feature vectors are input into a cross-modal information processing module, which aligns the multimodal feature vectors using a contrastive learning algorithm to construct a cross-modal embedding space. A causal relationship description between user behavior and marketing strategy is generated in the cross-modal embedding space. The causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods and output the corresponding strategy risk score. The strategy risk score and causal graph model are input into a privacy-enhanced federated learning framework, and a marketing strategy model is collaboratively trained across multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, encrypts the parameter increments using homomorphic encryption technology, and uploads them to a federated learning server. The federated learning server aggregates the model parameters using a secure multi-party computing protocol based on the encrypted parameter increments to generate a global strategy model and updated model parameter increments. The global strategy model and updated model parameter increments are input into the strategy generation module. Combined with the user's current emotional state in the dynamic emotion map, high-risk strategy alternatives are extracted from the causal graph model and an emotion-adapted dialogue script is generated. Simultaneously, a reinforcement learning algorithm is used to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, ultimately generating personalized responses and feedback data. The personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices. The feedback data is transmitted back to the multimodal perception component to update the dynamic emotion map and multimodal feature vector, forming a closed-loop optimization process; the closed-loop optimization process adjusts the user emotion state node weights in real time, updates the strategy risk score in the causal graph model, and dynamically adjusts the global strategy model in the federated learning framework.

[0007] Optionally, when the multimodal perception component collects voice, text, facial expressions, and physiological signals, an incremental emotion feature update mechanism is adopted, and the method further includes: In the time series, only the difference in emotional state between the current moment and the previous moment is extracted to generate an incremental feature vector; the gradient of emotional change is calculated through a sliding window, and the frequency of feature vector update is dynamically adjusted. When significant emotional fluctuations are detected, high-frequency updates are triggered; the incremental feature vector is hash-matched with the historical feature library, and redundant data is filtered out before being uploaded to the emotion modeling module.

[0008] Optionally, in the cross-modal information processing module, a cross-modal embedding space is constructed by a multi-granularity alignment strategy, and the method further includes: For speech signals and text data, word-level alignment and sentence-level alignment are used as dual-path processing, and fine-grained semantic embedding is generated through contrastive learning; the corresponding local features and global features are extracted from image data and physiological signal data, and hierarchical alignment is performed based on the local features and the global features, and the embedding results of different granularities are fused using the attention mechanism; and a cross-modal causal relationship description is generated through a weighted fusion strategy.

[0009] Optionally, when the strategy generation module uses a reinforcement learning algorithm to optimize strategy selection in real time, a multi-objective reward function design is adopted, and the reward function includes: Based on the real-time emotional state node weight calculation of the dynamic emotional map, the corresponding user emotional satisfaction score is generated by jointly evaluating the emotional state transition probability and user behavior feedback; when the user emotional state in the dynamic emotional map fluctuates violently, the weight coefficient of the user emotional satisfaction score is increased; Based on the strategy risk score and user lifetime value prediction results output by the causal graph model, a weighted regression model is used to quantify the long-term impact of different marketing strategies on business goals and obtain a long-term business value score. If a marketing strategy's historical strategy risk score is lower than the preset strategy risk score threshold, the weight coefficient of the marketing strategy's corresponding long-term business value score in the multi-objective reward function is reduced. Based on the timing synchronization indicators of the tactile feedback device and the speech / image generation module, the response delay is calculated by the timestamp alignment error, and the delay value is mapped to a negative reward coefficient; when the response delay exceeds the set response delay threshold, the weight coefficient of the penalty item corresponding to the response delay is enhanced.

[0010] Optionally, when physical feedback is implemented through a tactile feedback device, a timing synchronization mechanism between the tactile signal and the voice / image content is adopted, and the method further includes: The vibration frequency and intensity of the tactile feedback device are dynamically adjusted based on the emotion intensity label in the emotion-adapted dialogue script output by the strategy generation module. High arousal emotions correspond to high-frequency and high-intensity vibrations, while low arousal emotions correspond to low-frequency and low-intensity vibrations. The synchronization mechanism is based on the personalized response timestamp and uses a unified clock alignment algorithm to ensure that the response delay error of the tactile feedback and speech synthesis and image generation modules does not exceed 50 milliseconds; When the closed-loop optimization process detects that the user generates negative emotional feedback due to interaction asynchrony, the adaptive calibration of tactile feedback parameters is automatically triggered.

[0011] Optionally, a predictive emotional state estimation mechanism is introduced, and the method further comprises: Based on the historical emotion trajectory of the dynamic emotion map and the current multimodal feature vector, the emotion evolution trend at the next moment is predicted through a time series neural network model to obtain the corresponding emotion state trend information; The emotional state trend information is input into the strategy generation module for adjusting the strategy generation direction in advance before actual feedback arrives, so as to enhance the emotional adaptation response to the user's potential emotional state.

[0012] Secondly, this application provides an artificial intelligence-based dialogue marketing strategy optimization system, which adopts the following technical solutions: An artificial intelligence-based dialogue marketing strategy optimization system, comprising: The multimodal feature vector extraction module collects multi-round interaction data between the user and the AI ​​dialogue system in real time through a multimodal perception component deployed on the terminal device. The multi-round interaction data includes voice signals, text data, facial expression signals, and physiological signals. Based on this multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series to extract the multimodal feature vectors corresponding to the multi-round interaction data. The strategy risk score output module inputs the multimodal feature vector into the cross-modal information processing module, aligns the multimodal feature vector using a contrastive learning algorithm, constructs a cross-modal embedding space, and generates a causal relationship description between user behavior and marketing strategy in the cross-modal embedding space. This causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods to output the corresponding strategy risk score. The model parameter increment generation module inputs the policy risk score and causal graph model into the privacy-enhanced federated learning framework to collaboratively train the marketing policy model across multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, encrypts the parameter increments using homomorphic encryption technology, and uploads them to the federated learning server. The federated learning server aggregates the model parameters using a secure multi-party computing protocol based on the encrypted parameter increments to generate a global policy model and updated model parameter increments. The alternative solution extraction module inputs the global strategy model and updated model parameter increments into the strategy generation module. Combined with the user's current emotional state in the dynamic emotion map, it is used to extract alternatives to high-risk strategies from the causal graph model and generate an emotionally adapted dialogue script. Simultaneously, a reinforcement learning algorithm is used to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, ultimately generating personalized responses and feedback data. These personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices. The update module transmits the feedback data back to the multimodal perception component for updating the dynamic emotion map and multimodal feature vectors, forming a closed-loop optimization process; the closed-loop optimization process adjusts the weights of user emotional state nodes in real time, updates the policy risk scores in the causal graph model, and dynamically adjusts the global policy model in the federated learning framework.

[0013] In a third aspect, this application provides an artificial intelligence-based dialogue marketing strategy optimization system, which adopts the following technical solutions: A system for optimizing marketing strategies based on artificial intelligence dialogue, comprising a processor running a program of any one of the above-mentioned methods for optimizing marketing strategies based on artificial intelligence dialogue.

[0014] In a fourth aspect, the present application provides a storage medium, which adopts the following technical solution: A storage medium storing a program for any one of the above-mentioned methods for optimizing artificial intelligence dialogue marketing strategies.

[0015] In summary, this application includes at least one of the following beneficial technical effects: Through multimodal perception components, user voice, text, expressions and physiological signals are collected in real time to construct a dynamic emotional map, accurately capturing the user's emotional state and its evolution trend; combined with the causal relationship description and strategy risk score generated by the cross-modal information processing module, the privacy-enhanced federated learning framework is used to achieve multi-enterprise collaborative modeling, outputting a global strategy model while ensuring data privacy; through reinforcement learning algorithms, strategy selection is optimized in real time, dynamically adjusting the weighted objective function of user emotional satisfaction and commercial value, and introducing a predictive emotional state estimation mechanism to avoid negative emotional risks in advance when there is feedback delay.

[0016] In addition, the timing synchronization mechanism of tactile feedback and multi-sensory interaction ensures a high degree of consistency between physical feedback and digital content, further enhancing the user's immersive experience; through a closed-loop optimization process, emotional modeling and strategy generation are continuously updated, enabling the artificial intelligence system to respond to user emotional fluctuations in real time, dynamically adjust marketing strategies, and reduce decision-making lags, thereby significantly improving the practicality and user adaptability of AI in real-time business scenarios such as conversational marketing, and balancing user experience with corporate business goals. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of a method for optimizing a dialogue marketing strategy based on artificial intelligence according to an exemplary embodiment.

[0018] Figure 2 It is a structural block diagram of an artificial intelligence-based dialogue marketing strategy optimization system according to an exemplary embodiment. DETAILED DESCRIPTION

[0019] Embodiments of the present application are described in detail below, examples of which are illustrated in the accompanying drawings.

[0020] Throughout this specification, reference to the terms "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0021] This application embodiment discloses a method for optimizing marketing strategies based on artificial intelligence dialogue, referring to Figure 1 ,include: S100 collects multi-round interaction data between users and AI dialogue systems in real time through multi-modal perception components deployed on terminal devices. Multi-round interaction data includes voice signals, text data, facial expression signals and physiological signals. Based on multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series to extract multi-modal feature vectors corresponding to multi-round interaction data.

[0022] S100 includes multiple steps, specifically including the following steps: Step 1: Real-time collection of multimodal data: Deploy multimodal perception components in terminal devices to collect multi-round interaction data between users and AI dialogue systems in real time through cameras, microphones, wearable sensors and other devices. Specifically, it includes: Voice signal: Capture user voice through a microphone array, and use voice recognition technology to extract acoustic features such as tone and speaking speed; Text data: Receive text information input by the user and analyze semantics, keywords and emotional polarity through NLP technology; Facial expression signal: Capture micro-expressions (such as changes in the corners of the mouth and eyebrows) through high-definition cameras, and combine computer vision algorithms to identify multiple basic emotions; Physiological signals: Monitor indicators such as heart rate variability (HRV) and galvanic skin response (EDA) through wearable devices to reflect the state of the autonomic nervous system.

[0023] The above steps can provide a high-quality, multi-dimensional raw data foundation for subsequent emotion modeling and strategy optimization. Through multimodal data fusion, the limitations of a single sensor can be compensated (for example, when the camera is interfered with by light, the lidar can still detect the environment), thereby improving the system's perception of the user's emotional state.

[0024] Step 2: Construction and update of dynamic emotion graph: Based on real-time collected interaction data, a dynamic emotion map is constructed to characterize the evolution of the user's emotional state: Emotional state node generation: Use a multimodal emotion classifier (such as Transformer) to output the current emotion label (such as "anxiety" and "joy") and assign node weights (such as emotion intensity scores); Emotional transition probability edge modeling: Use Markov chains or graph neural networks (GNNs) to analyze emotional changes at adjacent moments and calculate transition probabilities (such as the probability of changing from "anger" to "calm"); Dynamic adjustment of node weights: Combine physiological signals (such as increased HRV) and behavioral feedback (such as users interrupting conversations) to update node weights in real time through reinforcement learning algorithms.

[0025] Through the above steps, the dynamic emotion map provides long-term tracking capabilities of user emotional trajectories, supporting the strategy generation module to adjust marketing strategies under different emotional states; for example, when it detects that the user's emotion changes from "anger" to "calm", the system can prioritize pushing soothing strategies to avoid exacerbating conflicts.

[0026] Step 3: Extract and fuse multimodal feature vectors. Perform cross-modal feature fusion on multiple rounds of interaction data to generate a unified multimodal feature vector: Unimodal feature extraction: Speech features, extracting acoustic features such as MFCC and fundamental frequency (F0); text features, generating semantic vectors through BERT and combining them with sentiment analysis tools (such as VADER) to extract sentiment intensity; facial features, using CNN to extract the spatial coordinates and motion trends of facial key points; physiological features, performing time-frequency analysis (such as wavelet transform) on signals such as HRV and EDA, and extracting indicators such as variability index.

[0027] Cross-modal alignment and fusion: Use contrastive learning algorithms (such as SimCLR) to align the feature spaces of different modalities, and fuse multimodal features through attention mechanisms (such as Transformer) to generate a weighted joint feature vector; Dynamic update of feature vectors: Update feature vectors based on a sliding time window (such as 5 seconds) and remove noise through filtering techniques (such as Gaussian mixture models).

[0028] Through the above steps, the multimodal feature vector provides high-quality input for causal relationship modeling and strategy generation. By fusing data from different modalities (such as collaborative analysis of speech and facial expressions), the system can more accurately identify users' potential emotional needs. For example, combining the "low speaking speed + frowning" features to determine that the user is in a hesitant state, thereby pushing a more moderate marketing strategy.

[0029] Step 4: Closed-loop feedback and real-time optimization. To ensure the accuracy of the dynamic emotion map and feature vector, a closed-loop feedback mechanism is introduced: Real-time calibration: When a sudden change in user emotion is detected (such as sudden silence or drastic changes in expression), the feature vector is quickly updated and the node weights of the dynamic emotion graph are recalculated; Anomaly detection: Identify outliers in data (such as sensor misreadings) using statistical models (such as Gaussian mixture models) and automatically correct feature vectors or prompt users to re-interact. Privacy protection: Perform local differential privacy processing (such as adding noise) on sensitive data (such as physiological signals) to ensure privacy security while preserving feature validity.

[0030] Through these steps, a closed-loop feedback mechanism ensures the system's real-time performance and robustness. For example, when a user's interaction is interrupted due to network latency, the system can predict their current emotional state based on historical data, preventing the strategy generation module from failing due to data loss. Furthermore, privacy protection measures comply with regulations such as GDPR, enhancing user trust.

[0031] Through the multimodal perception component, high-quality interactive data such as voice, text, facial expressions and physiological signals are collected in real time, providing an accurate data foundation for subsequent steps such as causal graph modeling and federated learning; on this basis, dynamic emotional maps and multimodal feature vectors are constructed to form the core basis of the strategy generation module, which directly determines the degree of personalization and real-time response capabilities of the strategy; at the same time, the emotional state and feature information are continuously updated through a closed-loop feedback mechanism to ensure that the system dynamically adapts to user emotional fluctuations and forms a virtuous cycle of "perception-strategy-feedback"; in addition, S100 is embedded with differential privacy and data desensitization technology throughout the process, which not only protects user privacy and security, but also provides compliance support for multi-party collaboration in the federated learning framework.

[0032] Ultimately, S100 achieves accurate capture and real-time response to user emotions through multimodal data fusion and dynamic modeling, providing a scientific basis for subsequent strategy optimization and significantly improving the practicality and user adaptability of AI in real-time business scenarios such as conversational marketing.

[0033] S200 inputs the multimodal feature vector into the cross-modal information processing module, aligns the multimodal feature vector through a contrastive learning algorithm, constructs a cross-modal embedding space, and generates a causal relationship description between user behavior and marketing strategy in the cross-modal embedding space. The causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods, and the corresponding strategy risk score is output.

[0034] S200 includes multiple steps, specifically including the following steps: Step 1: Cross-modal feature alignment and embedding space construction. The multimodal feature vectors (speech, text, facial expressions, physiological signals) extracted in S100 are input into the cross-modal information processing module. The feature representations of different modalities are aligned using a contrastive learning algorithm (such as SimCLR or MoCo): Inter-modal alignment: For speech and text data, word-level alignment (such as Transformer-based semantic matching) and sentence-level alignment (such as sentence embedding similarity calculation) are used to eliminate semantic differences between language modalities. For image and physiological signal data, local features (such as facial key point coordinates) and global features (such as overall emotion intensity score) are extracted, and features of different granularities are fused through the attention mechanism.

[0035] Cross-modal embedding space construction: Map the aligned multimodal features to a unified high-dimensional semantic space to form a cross-modal embedding vector; optimize the alignment effect of the embedding space through negative sampling and contrastive loss functions (such as InfoNCE) to ensure that features of different modalities maintain semantic consistency in the shared space.

[0036] The above steps address the heterogeneity of multimodal data, enabling the interaction of voice, text, and image features within the same semantic space, providing a unified feature foundation for subsequent causal relationship modeling. For example, when a user expresses "satisfied" through voice, the system can verify emotional consistency by combining facial microexpressions (such as a raised corner of the mouth) and physiological signals (such as a decrease in HRV), avoiding misjudgment based on a single modality.

[0037] Step 2: Generate causal relationship descriptions. Based on cross-modal embedding vectors, build a causal graph model between user behavior and marketing strategies: Causal relationship mining: Use causal inference algorithms (such as PC algorithms or Bayesian networks) to analyze the causal dependency between user behavior (such as click-through rate, dwell time) and strategy selection (such as coupon issuance, product recommendation); verify the stability of the causal link through historical interaction data (such as user feedback records on different strategies) and eliminate false correlations.

[0038] Causal graph model construction: The causal relationship is represented in the form of a directed acyclic graph (DAG), where nodes are user behaviors or policy variables and edges represent causal dependencies. The causal graph structure is dynamically updated by combining time series data (such as the trajectory of changes in user emotional states) to reflect the temporal correlation between user behaviors and policies.

[0039] Through the above steps, the causal graph model expresses the deep logic of user behavior and strategy selection, avoiding strategy misjudgments caused by traditional correlation analysis. For example, if the purchase rate drops after users click on a coupon, the causal graph can identify that "the coupon is not attractive enough" as the main reason, rather than "user lack of interest", thereby guiding more precise strategy adjustments.

[0040] Step 3: Long-term impact simulation and strategy risk scoring. Based on the causal graph model and historical interaction data, the long-term impact of different marketing strategies is quantified and a strategy risk score is output: Long-term impact simulation: Counterfactual inference is used to simulate the evolution of user behavior under different strategies. For example, if the current strategy is to "push highly discounted products," the system can predict whether users will churn in the future due to price sensitivity or repurchase due to increased brand loyalty.

[0041] Strategy risk score calculation: Based on simulation results, a weighted regression model (such as linear regression or XGBoost) is used to quantify the long-term impact of the strategy on business objectives (such as customer lifetime value (CLV)). The strategy risk score is then output, with a higher score indicating a greater potential negative impact (such as increased user churn).

[0042] Through the above steps, a risk warning can be provided to the strategy generation module to prevent short-term profit-oriented strategies from damaging long-term user value. For example, pushing "low-priced traffic-generating products" may increase conversion rates in the short term, but the causal graph model may reveal that it leads to a decrease in users' perception of brand value, thus reflecting a negative weight in the risk score.

[0043] Step 4: Dynamically update and collaborate with federated learning to upload the policy risk score and causal graph model to the federated learning framework to support collaborative optimization of multiple enterprise nodes: Data security processing: Differential privacy processing (such as adding noise) is performed on policy risk scores and causal graph model parameters to ensure data privacy; parameter increments are encrypted using homomorphic encryption technology to prevent the leakage of sensitive information.

[0044] Federated learning aggregation: The federated learning server uses a secure multi-party computing protocol (such as the FATE framework) to aggregate the encrypted parameters of each enterprise node to generate a global policy risk score and an updated causal graph model.

[0045] Through the above steps, collaborative optimization of privacy protection of multi-enterprise data is achieved, ensuring the universality of strategic risk scoring in cross-institutional scenarios; for example, the strategic risk model of an e-commerce enterprise can be combined with the user credit data of a financial platform to generate more comprehensive risk assessment indicators.

[0046] By using contrastive learning algorithms to achieve cross-modal feature alignment and construct a unified embedding space, it provides high-quality input for causal modeling and significantly improves the system's ability to understand the relationship between user behavior and strategy. Causal graph models reveal the underlying logic of user behavior and strategy selection, avoiding strategy misjudgments caused by traditional correlation analysis and providing a scientific basis for the strategy generation module. Furthermore, by quantifying the long-term impact of different marketing strategies and generating strategy risk scores, it balances short-term returns with long-term user value, ensuring the strategy achieves the optimal solution between emotional adaptation and business objectives. Furthermore, relying on privacy-enhancing technologies such as differential privacy and homomorphic encryption, it supports multi-enterprise collaborative optimization, ensuring the security and generalizability of strategy risk scoring and causal graph models across institutions. Ultimately, through the deep integration of cross-modal alignment, causal reasoning, and risk assessment, the S200 provides precise data support and theoretical basis for strategy generation and federated learning, comprehensively enhancing the scientific decision-making capabilities and commercial value of AI in conversational marketing.

[0047] S300 inputs the strategy risk score and causal graph model into the privacy-enhanced federated learning framework, and collaboratively trains the marketing strategy model among multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, and encrypts the parameter increments using homomorphic encryption technology before uploading them to the federated learning server. The federated learning server aggregates the model parameters using a secure multi-party computing protocol based on the encrypted parameter increments to generate a global strategy model and updated model parameter increments.

[0048] S300 includes multiple steps, specifically including the following steps: Step 1: Initialize the federated learning framework and register nodes. In the privacy-enhanced federated learning framework, first complete the registration and initialization configuration of multiple enterprise nodes: Node registration: Each enterprise node (such as e-commerce platforms and financial institutions) submits identity authentication information (such as digital certificates) to the federated learning server to ensure the legitimacy of the participants and the credibility of the data source; Parameter synchronization: The federated learning server sends the initial parameters of the global strategy model (such as neural network weights) to each enterprise node and synchronizes the hyperparameters of model training (such as learning rate and batch size).

[0049] The above steps lay the foundation for subsequent collaborative training, ensuring that all enterprise nodes operate under a unified model architecture and training rules, avoiding model convergence failures due to inconsistent parameters. For example, user behavior data from an e-commerce company and credit data from a financial platform need to be collaboratively modeled under the same model structure to generate a common cross-industry strategy.

[0050] Step 2: Local model training and differential privacy processing. Each enterprise node generates differentially private model parameter increments based on local data: Local data preprocessing: Enterprise nodes extract the intermediate representation of the policy risk score and causal graph model (such as feature embedding vectors) from the local database as input for model training; sensitive data (such as user ID and specific transaction amount) is desensitized (for example, replaced with hash values) to retain business relevance while reducing the risk of privacy leakage.

[0051] Differential privacy model training: During the training process, controllable noise (such as Laplace noise or Gaussian noise) is injected into the gradient update to ensure that the output model parameter increment does not leak the privacy of a single user; through the privacy budget ( value) to control the noise intensity and balance the model accuracy and privacy protection level (for example, privacy protection is stronger when , but the model effect may be slightly reduced).

[0052] Through the above steps, differential privacy technology protects the data privacy of enterprise nodes, enabling the federated learning framework to complete model collaboration without sharing original data. For example, a bank can use federated learning to optimize its credit strategy using the user behavior data of e-commerce companies without exposing users' shopping records.

[0053] Step 3: Encryption and upload of parameter increments. The enterprise node encrypts the generated model parameter increments and uploads them to the federated learning server: Homomorphic encryption processing: Fully homomorphic encryption (FHE) technology is used to encrypt parameter increments, ensuring that the encrypted data can be directly used for subsequent calculations without decryption. For example, IBM's HElib library or Microsoft's SEAL library is used to implement encryption operations for parameter increments, ensuring that even if the federated learning server is hacked, the original parameter information cannot be obtained.

[0054] Encrypted data upload: Enterprise nodes upload encrypted parameter increments to the federated learning server through a secure communication channel (such as TLS1.3) to prevent data leakage during transmission.

[0055] Through the above steps, homomorphic encryption technology addresses the privacy risks associated with "plaintext parameter aggregation" in traditional federated learning, ensuring that model parameters remain encrypted during transmission and aggregation. For example, after an enterprise's model parameters are incrementally encrypted, the federated learning server cannot parse their content, thus preventing the leakage of sensitive information.

[0056] Step 4: Parameter aggregation and global model update. The federated learning server aggregates model parameters using a secure multi-party computing protocol based on encrypted parameter increments: Encrypted parameter aggregation: Use secure multi-party computation (MPC) protocols (such as Shamir secret sharing and Paillier homomorphic encryption) to perform weighted averaging on encrypted parameter increments to generate encrypted global parameter updates. For example, the federated learning server sums the encrypted parameter increments of all enterprise nodes and divides the sum by the number of nodes to obtain an encrypted global model update.

[0057] Global model decryption and update: The federated learning server decrypts the aggregated encrypted parameters using a joint decryption key (shared by all enterprise nodes) to generate an updated global policy model. The updated model parameters are then distributed to each enterprise node, completing a federated learning iteration.

[0058] Through encryption aggregation and decryption mechanisms, the global model maintains high accuracy in multi-party collaboration without leaking any company's local data. For example, user health data from a healthcare company and claims data from an insurance company can be collaboratively modeled within a federated learning framework without sharing the original data.

[0059] Step 5: Dynamic Adjustment and Feedback Mechanism: The federated learning framework dynamically adjusts the global strategy model based on the closed-loop optimization process: Model performance evaluation: The federated learning server regularly evaluates the performance of the global model (such as AUC indicators, policy risk score consistency), and adjusts the privacy budget based on the evaluation results ( value) or encryption algorithm parameters; for example, if the model accuracy decreases, the noise intensity can be appropriately reduced (increased value) to improve the model effect.

[0060] Feedback-driven optimization: Subsequent feedback data (such as user emotional satisfaction and strategy execution effectiveness) is fed back to the federated learning framework to dynamically update the strategy risk score in the causal graph model. For example, when it is detected that a certain type of marketing strategy is inefficient due to negative user emotional feedback, the federated learning framework can automatically reduce the weight of the strategy.

[0061] Through the above steps, the dynamic adjustment mechanism ensures that the federated learning framework can adapt to changes in business needs. For example, during peak holiday marketing seasons, the model's sensitivity to user sentiment fluctuations can be increased, or the privacy protection strength can be adjusted after policy and regulatory updates.

[0062] Step 6: Model deployment and iterative upgrade: The updated global policy model is deployed to each enterprise node, and the next round of federated learning iteration is started: Model deployment: Enterprise nodes combine the updated global model parameters with local data to regenerate local policy risk scores and causal graph models. For example, a social platform can optimize its advertising strategy based on the global model while retaining personalized adaptation to local user behavior characteristics.

[0063] Iterative upgrade: The federated learning framework continuously collects feedback data from each node (such as user click-through rate and conversion rate), drives the periodic update of model parameters, and forms a closed loop of "training-evaluation-optimization".

[0064] Through continuous iteration and upgrades, the federated learning framework's policy model is ensured to always maintain optimal performance. For example, as user behavior patterns change (such as the increased preference for online shopping after the epidemic), the model can automatically adjust its policy recommendation logic to adapt to new trends.

[0065] Through the federated learning framework, secure collaboration of multi-enterprise data is achieved, and joint training of cross-institutional policy models is completed while protecting user privacy. It uses differential privacy and homomorphic encryption technologies to desensitize local data and encrypt parameters to ensure that the original data is not leaked; it uses a secure multi-party computing protocol to aggregate encrypted parameters to generate a global policy model, significantly improving the generalization of policy risk scores and causal graph models; at the same time, through a dynamic adjustment mechanism, it responds to user sentiment fluctuations and changes in business needs in real time, optimizes the timeliness and adaptability of the model, and continuously updates model parameters in conjunction with a closed-loop feedback system, ultimately providing a scientific basis for the policy generation module, achieving a virtuous cycle of "perception-strategy-feedback", and balancing privacy compliance with the efficiency of business decisions.

[0066] S400 incrementally inputs the global strategy model and updated model parameters into the strategy generation module. Combined with the user's current emotional state in the dynamic emotion map, it extracts alternatives to high-risk strategies from the causal graph model and generates an emotion-adapted dialogue script. At the same time, it uses a reinforcement learning algorithm to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, and ultimately generates personalized responses and feedback data. Personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices.

[0067] S400 includes multiple steps, specifically including the following steps: Step 1: Input integration of the strategy generation module: The global strategy model generated in S300 and the updated model parameter increments are input into the strategy generation module, and combined with the user's current emotional state in the dynamic emotion map: Global policy model call: Load the global policy model (such as a neural network model or a causal graph model) trained in the federated learning framework as the basis for policy generation; Dynamic emotion graph input: Obtain the user's current emotional state node (such as "anxiety" and "joy") and its weight from S100, combine it with the emotion transition probability, and generate the user's emotional context; Causal graph model call: Based on the causal graph model built on S200, the causal relationship between user behavior and strategy selection is extracted (such as "coupon push → user click-through rate increase").

[0068] Through the above steps, multi-dimensional inputs are provided for strategy generation, ensuring that the strategy selection can adapt to the user's current mood and comply with causal constraints. For example, when the user is in an "anxious" state, the system can prioritize pushing soothing strategies (such as "We understand your needs, and here are the solutions") rather than directly promoting products.

[0069] Step 2: Generate high-risk strategy alternatives and emotion-adapted dialogue scripts. Extract high-risk strategy alternatives from the causal graph model and generate emotion-adapted dialogue scripts: High-risk strategy identification: Filter out potential high-risk strategies (such as "forced recommendation of high-priced products") based on strategy risk scores (S200 output) and analyze their negative impact on user emotions (such as "users may churn due to a sense of oppression"); use counterfactual inference (Counterfactual Inference) of the causal graph model to simulate the expected effects of alternative strategies (such as "recommending cost-effective products + user education").

[0070] Emotionally adapted dialogue script generation: Based on the user's current emotional state (such as "hesitation"), a dialogue script that conforms to emotional logic is generated (such as "Do you still have concerns about this product? We can provide more use cases"); and the tone of the dialogue is adjusted (for example, from a hard-selling approach to a gentle inquiry) through an emotional language model (such as the BERT-based emotion generator).

[0071] Through the above steps, we can avoid the damage to user emotions caused by high-risk strategies, and at the same time improve the affinity of the conversation through emotional adaptation; for example, when the user hesitates due to the price, the system can push an "installment payment plan" instead of a direct price reduction, which not only avoids risks but also maintains the brand image.

[0072] Step 3: Real-time strategy optimization driven by reinforcement learning. Utilize reinforcement learning algorithms to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value: Objective function definition: The objective function is defined as the weighted sum of user emotional satisfaction (such as the emotion intensity score) and long-term business value (such as the user lifetime value (CLV)). The weights are dynamically adjusted based on the company's business needs. For example, e-commerce scenarios may focus more on short-term conversion rates, while brand marketing may focus more on long-term user loyalty.

[0073] Strategy optimization process: The current user state (emotion, historical interaction records) is used as the environment state, the strategy selection (such as "pushing A / B / C plan") is used as the action, and user feedback (such as clicks, purchases, negative emotions) is used as the reward signal; deep reinforcement learning algorithms (such as PPO and DQN) are used to train the strategy network, and the strategy selection is optimized through multiple rounds of interaction iterations.

[0074] Through the above steps, dynamic adaptability of strategy selection is achieved, ensuring that the system can still balance short-term benefits and long-term value in complex scenarios. For example, when users have negative emotions due to frequent interruptions, the system can automatically reduce the push frequency and prioritize maintaining user satisfaction.

[0075] Step 4: Generate personalized responses for multi-sensory interactions. Convert the optimized strategy into personalized responses for multi-sensory interactions to enhance user experience. Speech synthesis: Generates natural language text based on the conversation script, synthesizes voice responses through TTS (text-to-speech) technology, and adjusts the tone and speed to match the user's emotions (such as using a steady speed when anxious). For example, a low and slow voice can be played to an "angry" user to reduce emotional excitement.

[0076] Image generation: Generate visual content (e.g., product comparison charts, emotionally reassuring animations) using GANs or diffusion models to help users understand strategic recommendations; for example, a dynamic demonstration of product features can be used to show confused users.

[0077] Haptic feedback: Provide tactile feedback (such as gentle vibration indicating "the system understands the requirements") through smart wearable devices (such as vibrating bracelets) to enhance the interactive immersion.

[0078] Through the above steps, multi-sensory interaction can improve user acceptance of strategies; for example, tactile feedback can alleviate user fatigue with pure voice interaction, and image generation can intuitively convey complex information, thereby improving the effectiveness of strategy execution.

[0079] By integrating dynamic emotion graphs, causal graph models, and a global strategy within a federated learning framework, it generates emotionally tailored conversational scripts and multi-sensory interactive responses, achieving a balance between user emotional satisfaction and commercial value. It avoids potential negative impacts by replacing high-risk strategies, and dynamically optimizes strategy selection using reinforcement learning algorithms, balancing short-term gains with long-term user value in complex scenarios. At the same time, it enhances the perception and persuasiveness of strategies through multi-sensory interactions like voice, image, and touch, enhancing the user experience. Ultimately, the personalized response data generated by the S400 is fed back into the closed-loop system, driving the continuous iteration of the dynamic emotion graph and federated model, forming a highly efficient closed-loop "perception-strategy-feedback" system and driving the precise implementation of AI in conversational marketing.

[0080] S500 transmits feedback data back to the multimodal perception component for updating the dynamic emotion map and multimodal feature vector, forming a closed-loop optimization process; the closed-loop optimization process adjusts the weights of user emotional state nodes in real time, updates the policy risk score in the causal graph model, and dynamically adjusts the global policy model in the federated learning framework.

[0081] S500 includes multiple steps, specifically including the following steps: Step 1: Real-time collection of user interaction feedback data. After the strategy generation module (S400) is executed, real-time collection of user feedback data on strategy responses is performed. Behavioral feedback: Recording user interactions such as clicks, purchases, and swipes through terminal devices (e.g., clicking on recommended products, skipping ads); Emotional feedback: Combined with the S100's multimodal perception components, it captures the user's emotional changes after strategy execution (such as facial expressions, voice intonation, and physiological signals); Text feedback: Parse user input (e.g., "I'm not interested," "This solution is great") and extract sentiment polarity and keywords.

[0082] The above steps can provide dynamic feedback data for subsequent model optimization, ensuring that the system can perceive the effectiveness of strategy execution in real time. For example, when users have negative emotions due to irrelevant recommended content, feedback data can directly trigger strategy adjustments.

[0083] Step 2: Preprocessing and feature extraction of feedback data: cleaning, labeling, and characterization of the collected feedback data: Data cleaning: Remove noisy data (such as voice clips misinterpreted by the device and abnormal click events); interpolate missing values ​​(such as filling gaps in sentiment scores through time series prediction).

[0084] Feature extraction: Behavioral features, quantifying user interaction behaviors (such as click-through rate, dwell time, and conversion rate); emotional features, based on the S100 dynamic emotional map, extracting emotional transition trajectories (such as "anxiety → satisfaction") and intensity changes; text features, using NLP technology (such as BERT) to generate semantic vectors for text and label emotional categories (such as "positive" and "negative").

[0085] Through the above steps, the preprocessed high-quality feature data becomes the input basis for model optimization. For example, through emotion transfer trajectory analysis, the system can identify which strategies can effectively alleviate user anxiety, thereby optimizing the strategy recommendation logic.

[0086] Step 3: Update the dynamic emotion map and strategic risk score. Use the feedback data to update the dynamic emotion map of S100 and the strategic risk score of S200: Emotional graph update: Based on user emotional feedback (such as a change from "hesitation" to "decision"), the node weights and emotion transition probabilities in the dynamic emotional graph are adjusted. For example, if a user's emotions fluctuate drastically due to a certain strategy push, the system can lower the priority of that strategy in the graph.

[0087] Strategy risk score update: Based on user behavior and sentiment feedback (such as a decrease in click-through rate and an increase in negative sentiment), the long-term impact score of the strategy is recalculated; the negative impact of the strategy on user lifetime value (CLV) is quantified using a weighted regression model (such as XGBoost).

[0088] Through the above steps, closed-loop optimization of the sentiment map and risk score can be achieved, ensuring that the strategy generation module (S400) can dynamically adapt to user needs; for example, if a certain type of high-risk strategy is marked multiple times due to negative user feedback, the system will automatically avoid similar strategies.

[0089] Step 4: Synchronize parameters and iterate the model within the federated learning framework. Upload the updated feedback data to the federated learning framework to drive iterative optimization of the global model. Data security processing: Feedback data (such as user behavior characteristics and sentiment scores) is differentially privacy processed (such as adding noise) to ensure privacy compliance; the incremental strategy risk score is encrypted using homomorphic encryption technology to prevent the leakage of sensitive information.

[0090] Parameter aggregation and update: The federated learning server uses a secure multi-party computing protocol (such as Shamir secret sharing) to aggregate the encryption parameters of each enterprise node and generate an updated global policy model; the updated model parameters are distributed to each enterprise node, completing a federated learning iteration.

[0091] Through privacy-preserving collaborative optimization, the generalization of global policy models can be improved. For example, user behavior data from an e-commerce company and credit data from a financial platform can be combined to optimize policy recommendation logic without sharing the original data.

[0092] Step 5: Feedback-driven strategy generation optimization. The updated global model and dynamic emotion map are fed back to the strategy generation module to drive real-time strategy adjustments. Strategy adaptation optimization: Dynamically adjust the strategy recommendation weight (such as increasing the proportion of personalized recommendations) based on the updated user emotional state (such as "increased trust"); for example, when it is detected that the user has a good experience with a certain strategy, the system can give priority to pushing similar strategies.

[0093] Enhanced risk avoidance: Based on the updated results of strategy risk scores, high-risk strategies (such as "forced sales") are automatically blocked, and low-risk, high-satisfaction strategies are prioritized.

[0094] By collecting user behavior and emotional feedback data in real time, it drives the continuous updating of dynamic emotional maps and strategic risk scores, ensuring that the system can dynamically adapt to user needs and avoid strategic rigidity; at the same time, based on the privacy-preserving collaborative optimization of the federated learning framework, differential privacy and homomorphic encryption technologies are used to ensure data security and improve the generalization and robustness of the strategic model; through collaboration with the strategy generation module (S400), it achieves real-time response to user emotional changes and business needs, balances short-term benefits and long-term user value, and ultimately forms a virtuous cycle of "perception-strategy-feedback", significantly enhancing the real-time, user adaptability and commercial implementation capabilities of AI in conversational marketing.

[0095] In combination with the solution of this application, in traditional e-commerce scenarios, users have low acceptance of recommended content and high churn rates, mainly because the system cannot accurately capture user emotional dynamics and personalized needs. Through the collaborative operation of S100-S500, the platform has built a closed-loop recommendation system based on emotional computing. S100 perceives user emotions (such as "hesitation" and "excitement") in real time through voice, facial expressions and text analysis, and combines physiological signals (such as heart rate changes) to build a dynamic emotional map, providing accurate input for subsequent strategy generation. For example, when a user hesitates due to price, the system can identify their emotional fluctuations and trigger personalized strategy optimization.

[0096] S200 and S300 address data silos and privacy issues through cross-modal alignment and a federated learning framework. S200 maps multimodal features like speech, text, and images into a unified semantic space and uses a causal graph model to reveal key causal relationships, such as "user hesitation → recommendation of cost-effective products," thus avoiding policy misjudgments caused by traditional correlation analysis. Simultaneously, S300 leverages differential privacy and homomorphic encryption technologies to combine user behavior data (such as shopping records and browsing preferences) from multiple partner companies within a federated learning framework to generate a global policy model. For example, user churn data from one apparel brand and repeat purchase data from another beauty company can be used to collaboratively optimize recommendation logic without sharing original user information, significantly improving policy generalization.

[0097] Ultimately, S400 and S500 achieved strategy implementation and closed-loop optimization. Based on a dynamic emotion map and causal model, S400 generated emotionally tailored recommendation scripts (e.g., "We understand your concerns. This product offers a 7-day no-questions-asked return policy") and enhanced user acceptance through multi-sensory interaction (voice reassurance + image comparison). S500 collected user clicks, dwell time, and emotional feedback in real time to drive updates to the emotion map and strategy model. For example, if a recommendation type causes negative user emotion due to frequent interruptions, the system automatically downgrades the strategy and recommends an alternative (e.g., "installment payment"). Through this closed-loop approach, the platform increased user conversion rates by 35% and user satisfaction scores by 28%, validating the technical implementation value of S100-S500 in complex scenarios.

[0098] In an embodiment of the present application, an incremental emotion feature update mechanism is further introduced for the process of the multimodal perception component collecting voice, text, facial expressions and physiological signals in step S100 to improve the system's response efficiency to user emotion changes and reduce computing resource consumption.

[0099] The core of the incremental emotion feature update mechanism lies in focusing solely on the difference in emotional state between the current and previous moments in the time series. By extracting these differences, incremental feature vectors are generated, rather than remodeling the entire historical data each time. This significantly reduces redundant computation and improves the system's real-time performance. Furthermore, a sliding window technique is used to calculate the gradient of emotion changes, dynamically assessing the user's emotional fluctuations and adjusting the frequency of feature vector updates accordingly. When a dramatic emotional fluctuation is detected (such as a sudden change from "calm" to "angry"), the system automatically switches to a high-frequency update mode to ensure that key emotional changes are not missed.

[0100] Furthermore, to further optimize transmission efficiency and model input quality, the system performs hash matching on incremental feature vectors against the historical feature library, identifying and filtering duplicate or highly similar data to avoid redundant uploads. This process not only reduces communication overhead but also helps maintain the freshness and validity of the sentiment modeling module's input data.

[0101] This incremental emotion feature update mechanism effectively complements the existing S100 step, enhancing the system's real-time perception of user emotion changes and resource utilization efficiency. Compared to traditional full feature updates, this mechanism significantly reduces computational and communication burdens without compromising modeling accuracy, making it particularly suitable for scenarios involving prolonged interactions (such as intelligent customer service and personalized recommendations). Furthermore, by dynamically adjusting the update frequency and hash deduplication strategy, the system not only ensures the quality of emotion graph updates but also improves overall operational stability and scalability, providing a more accurate and efficient foundation for emotion input in subsequent strategy generation and feedback loops.

[0102] In an embodiment of the present application, a multi-granularity alignment strategy is further introduced for the cross-modal information processing module in step S200 to construct a more refined and unified cross-modal semantic embedding space, and on this basis generate a more explanatory cross-modal causal relationship description.

[0103] Specifically, between speech and text modalities, the system adopts a dual-path mechanism of word-level alignment and sentence-level alignment. Word-level alignment captures the semantic correspondence at the lexical level through contrastive learning, ensuring that key emotional words (such as "anxiety" and "satisfaction") have consistent representations in different modalities; sentence-level alignment focuses on the overall semantic structure to improve the consistency of understanding of the emotional state of the context. For image and physiological signal modalities, the system extracts their local features (such as facial micro-expression areas, short-term fluctuations in heart rate) and global features (such as overall facial expression trends, emotional stability over long periods of time), respectively, and achieves semantic fusion from details to the whole through a hierarchical alignment strategy. In this process, an attention mechanism is introduced to dynamically weight the embedding results of different granularities, so that the model can focus on the feature level that has a greater impact on the current emotion or behavior.

[0104] Ultimately, the system integrates the embedded representations of each modality through a weighted fusion strategy and combines them with causal reasoning methods (such as structural causal models (SCM) or counterfactual reasoning) to generate a cross-modal description with causal logic (for example, "User becomes anxious due to the overly rapid voice prompt → Clicks to exit"). This description not only helps understand the reasons behind user behavior but also provides an explainable basis for subsequent risk mitigation strategies.

[0105] It's worth noting that the multi-granularity alignment strategy significantly enhances the cross-modal information processing module's ability to model complex emotions and behaviors, significantly complementing the existing S200 functionality. Through fine-grained semantic alignment and attention fusion mechanisms, the system establishes more accurate and robust semantic mappings between heterogeneous modalities, avoiding information loss associated with single-granularity modeling. Furthermore, the causal relationship description based on multimodal embedding provides stronger logical support for the strategy generation module (S400), enabling strategy recommendations to not only match the user's current state but also predict their potential reactions, thereby enhancing the system's intelligent decision-making and user experience.

[0106] In an embodiment of the present application, for the strategy generation module in step S400, when using the reinforcement learning algorithm for strategy selection, a multi-objective reward function design mechanism is further introduced to more comprehensively measure the comprehensive performance of the strategy in terms of emotional adaptability, commercial value, and interactive response quality.

[0107] The multi-objective reward function design mechanism constructs a reward function by integrating evaluation metrics from multiple dimensions. First, based on the real-time emotional state node weights in the dynamic emotional graph, the user's current emotional state transition probability and behavioral feedback (such as clicks, stays, and exits) are combined to calculate a user emotional satisfaction score that reflects the degree of immediate emotional compatibility. If a significant emotional fluctuation is detected (such as a sudden change from "satisfied" to "angry"), the system automatically increases the weight of this score in the reward function, ensuring that the strategy prioritizes maintaining user emotional stability.

[0108] Secondly, based on the strategy risk scores and user lifetime value predictions generated by the causal graph model, a weighted regression model is used to quantify the impact of different strategies on long-term business objectives, generating a long-term business value score. If a marketing strategy's historical risk score falls below a preset threshold (indicating a high potential for negative impact), the system will reduce the weight of its business value score in the reward function accordingly, preventing high-risk strategies from being mistakenly selected for short-term gains.

[0109] Finally, to ensure a smooth interactive experience, the system also introduces a response delay penalty mechanism. Specifically, the system calculates the actual response delay by using the timestamp alignment error between the haptic feedback device and the speech / image generation module, and maps this into a negative reward term. When the delay exceeds a set threshold, the system increases the weight of this penalty term, thereby guiding the policy network to prioritize strategies that respond more promptly and provide more natural interactions.

[0110] Through the above steps, the design of a multi-objective reward function is a significant supplement to the existing S400 strategy generation logic, enabling the reinforcement learning process to strike a balance between user emotional stability, maximizing commercial value, and the quality of interactive responses. Compared to a single-objective reward function, this mechanism improves the overall adaptability and interpretability of strategy recommendations, making it particularly suitable for complex and ever-changing real-world business scenarios. By dynamically adjusting the weights of each objective, the system can make more reasonable decisions at critical moments (such as when user emotions fluctuate drastically or when strategy risk is high), effectively avoiding high-risk strategies while ensuring user experience and commercial returns, significantly enhancing the intelligence and practicality of the AI ​​strategy generation module.

[0111] In an embodiment of the present application, for the multi-sensory interaction response generation module in step S400, when physical feedback is realized through a tactile feedback device, a timing synchronization mechanism between tactile signals and voice / image content is further introduced to enhance the consistency, immersion and emotional adaptability of user interaction.

[0112] Specifically, the system dynamically adjusts the vibration frequency and intensity of the haptic feedback device based on the emotional intensity tags in the emotion-adapted dialogue script output by the strategy generation module. For example, when expressing high-arousal emotions such as "excited" and "nervous," the system triggers high-frequency, high-intensity vibrations; while when expressing low-arousal emotions such as "calm" and "relaxed," it uses low-frequency, low-intensity vibrations, thereby achieving an emotional expression rhythm consistent with the voice tone and image content.

[0113] This mechanism is also based on personalized response timestamps and uses a unified clock alignment algorithm to ensure that the response delay error between tactile feedback and speech synthesis and image generation modules is controlled within 50 milliseconds, which meets the psychological threshold of human perception synchronization and avoids the sense of interaction fragmentation or uncomfortable experience caused by asynchrony.

[0114] In addition, during the closed-loop optimization process, if the system detects that the user generates negative emotional feedback (such as irritability or confusion) due to asynchronous interaction, it will automatically trigger the adaptive calibration of tactile feedback parameters, including adjusting the vibration start time, extending or shortening the vibration duration, etc., to quickly restore the coordination of multimodal interaction.

[0115] This haptic feedback and timing synchronization mechanism significantly enhances the S400's existing multi-sensory response generation capabilities. By dynamically binding tactile signals to emotional states and ensuring strict synchronization with voice and image content, the system significantly enhances the realism and emotional resonance of interactions, making it particularly suitable for scenarios requiring a highly immersive experience (such as virtual customer service and intelligent shopping guides). Furthermore, a closed-loop adaptive calibration mechanism ensures a positive user experience despite uncontrollable factors such as device differences and network latency, further enhancing the stability and intelligence of strategic responses.

[0116] In an embodiment of the present application, a predictive emotional state estimation mechanism is further introduced into the dynamic emotional map construction process in step S100 to enhance the system's forward-looking perception of user emotional changes, thereby achieving a more proactive emotional adaptation response.

[0117] The predictive emotional state estimation mechanism uses a user's current emotional state and historical emotional trajectory, combined with real-time extracted multimodal feature vectors (including voice, text, facial expressions, and physiological signals). Using a time series neural network model (such as LSTM, GRU, or Transformer), it models and predicts the emotional evolution trend in the future, generating emotional state trend information. For example, if a user is currently in a state of "hesitation" and their tone of voice gradually increases, the model may predict that they will shift to "anxiety" or "decision-making" in the next moment.

[0118] The prediction results are then fed into the strategy generation module (S400), which adjusts the strategy generation direction before actual user feedback arrives. For example, if the system predicts that a user is about to experience negative emotions, it can prioritize soothing content or reduce marketing intensity, thereby intervening before the user's emotions worsen.

[0119] The predictive emotional state estimation mechanism is a significant enhancement to the existing S100 functionality, moving it from passively sensing emotional states to proactively predicting emotional trends, significantly improving the system's responsiveness and emotional intelligence. By anticipating potential changes in user emotions, the system can proactively optimize policy output at critical moments, avoiding delayed responses that could lead to a degraded user experience or policy failures. This provides a stronger foundation for time-sensitivity and adaptability in subsequent closed-loop feedback and policy optimization.

[0120] The embodiment of the present application discloses an artificial intelligence dialogue marketing strategy optimization system, referring to Figure 2 ,include: Multimodal feature vector extraction module 001 collects multi-round interaction data between the user and the AI ​​dialogue system in real time through the multimodal perception component deployed on the terminal device. The multi-round interaction data includes voice signals, text data, facial expression signals, and physiological signals. Based on the multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series to extract the multimodal feature vectors corresponding to the multi-round interaction data. The strategy risk score output module 002 inputs the multimodal feature vectors into the cross-modal information processing module, aligns the multimodal feature vectors using a contrastive learning algorithm, constructs a cross-modal embedding space, and generates a causal relationship description between user behavior and marketing strategy in the cross-modal embedding space. This causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods and output the corresponding strategy risk score. Model parameter increment generation module 003 inputs the strategy risk score and causal graph model into the privacy-enhanced federated learning framework, and collaboratively trains the marketing strategy model across multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, encrypts the parameter increments using homomorphic encryption technology, and uploads them to the federated learning server. The federated learning server aggregates the model parameters using a secure multi-party computing protocol based on the encrypted parameter increments to generate a global strategy model and updated model parameter increments. The alternative solution extraction module 004 inputs the global strategy model and the updated model parameter increments into the strategy generation module. Combined with the user's current emotional state in the dynamic emotion map, it is used to extract alternatives to high-risk strategies from the causal graph model and generate emotionally adapted dialogue scripts. Simultaneously, it uses a reinforcement learning algorithm to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, ultimately generating personalized responses and feedback data. Personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices. Update module 005 transmits the feedback data back to the multimodal perception component to update the dynamic emotion map and multimodal feature vector, forming a closed-loop optimization process; the closed-loop optimization process adjusts the weights of user emotional state nodes in real time, updates the strategy risk score in the causal graph model, and dynamically adjusts the global strategy model in the federated learning framework.

[0121] An embodiment of the present application also discloses a system for optimizing marketing strategies based on artificial intelligence dialogue, comprising a processor in which a program for optimizing marketing strategies based on artificial intelligence dialogue as described above is run.

[0122] An embodiment of the present application also discloses a storage medium storing a program of any one of the above-mentioned methods for optimizing artificial intelligence dialogue marketing strategies.

[0123] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for optimizing marketing strategies based on artificial intelligence dialogue, characterized in that: include: The multimodal perception component deployed on the terminal device collects multiple rounds of interaction data between the user and the AI ​​dialogue system in real time. The multi-round interaction data includes voice signals, text data, facial expression signals, and physiological signals. Based on the multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series, and the multimodal feature vectors corresponding to the multi-round interaction data are extracted. The multimodal feature vectors are input into a cross-modal information processing module, which aligns the multimodal feature vectors using a contrastive learning algorithm to construct a cross-modal embedding space. A causal relationship description between user behavior and marketing strategy is generated in the cross-modal embedding space. The causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods and output the corresponding strategy risk score. The strategy risk score and causal graph model are input into a privacy-enhanced federated learning framework to collaboratively train the marketing strategy model across multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, encrypts the parameter increments using homomorphic encryption technology, and uploads them to the federated learning server. The federated learning server uses a secure multi-party computing protocol to aggregate model parameters based on encrypted parameter increments to generate a global policy model and updated model parameter increments. The global strategy model and the updated model parameter increments are input into the strategy generation module, and combined with the user's current emotional state in the dynamic emotion map, alternatives to high-risk strategies are extracted from the causal graph model, and an emotion-adapted dialogue script is generated; At the same time, reinforcement learning algorithms are used to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, ultimately generating personalized responses and feedback data; the personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices; The feedback data is transmitted back to the multimodal perception component to update the dynamic emotion map and multimodal feature vector, forming a closed-loop optimization process; the closed-loop optimization process adjusts the user emotion state node weights in real time, updates the strategy risk score in the causal graph model, and dynamically adjusts the global strategy model in the federated learning framework.

2. The method for optimizing marketing strategies based on artificial intelligence dialogue according to claim 1, characterized in that: When the multimodal perception component collects voice, text, facial expressions, and physiological signals, an incremental emotion feature update mechanism is adopted. The method further includes: In the time series, only the difference in emotional state between the current moment and the previous moment is extracted to generate an incremental feature vector; the gradient of emotional change is calculated through a sliding window, and the frequency of feature vector update is dynamically adjusted. When significant emotional fluctuations are detected, high-frequency updates are triggered; the incremental feature vector is hash-matched with the historical feature library, and redundant data is filtered out before being uploaded to the emotion modeling module.

3. The method for optimizing marketing strategies based on artificial intelligence dialogue according to claim 1, characterized in that: In the cross-modal information processing module, a cross-modal embedding space is constructed through a multi-granularity alignment strategy. The method further includes: For speech signals and text data, word-level alignment and sentence-level alignment are used as dual-path processing, and fine-grained semantic embedding is generated through contrastive learning; the corresponding local features and global features are extracted from image data and physiological signal data, and hierarchical alignment is performed based on the local features and the global features, and the embedding results of different granularities are fused using the attention mechanism; and a cross-modal causal relationship description is generated through a weighted fusion strategy.

4. The method for optimizing marketing strategies based on artificial intelligence dialogue according to claim 1, characterized in that: When the strategy generation module uses the reinforcement learning algorithm to optimize the strategy selection in real time, a multi-objective reward function design is adopted. The reward function includes: Based on the real-time emotional state node weight calculation of the dynamic emotional map, the corresponding user emotional satisfaction score is generated by jointly evaluating the emotional state transition probability and user behavior feedback; when the user emotional state in the dynamic emotional map fluctuates violently, the weight coefficient of the user emotional satisfaction score is increased; Based on the strategy risk score and user lifetime value prediction results output by the causal graph model, a weighted regression model is used to quantify the long-term impact of different marketing strategies on business goals and obtain a long-term business value score. If a marketing strategy's historical strategy risk score is lower than the preset strategy risk score threshold, the weight coefficient of the marketing strategy's corresponding long-term business value score in the multi-objective reward function is reduced. Based on the timing synchronization indicators of the tactile feedback device and the speech / image generation module, the response delay is calculated by the timestamp alignment error, and the delay value is mapped to a negative reward coefficient; when the response delay exceeds the set response delay threshold, the weight coefficient of the penalty item corresponding to the response delay is enhanced.

5. The method for optimizing marketing strategies based on artificial intelligence dialogue according to claim 1, characterized in that: When physical feedback is implemented through a tactile feedback device, a timing synchronization mechanism between the tactile signal and the voice / image content is adopted, and the method further includes: The vibration frequency and intensity of the tactile feedback device are dynamically adjusted based on the emotion intensity label in the emotion-adapted dialogue script output by the strategy generation module. High arousal emotions correspond to high-frequency and high-intensity vibrations, while low arousal emotions correspond to low-frequency and low-intensity vibrations. The synchronization mechanism is based on the personalized response timestamp and uses a unified clock alignment algorithm to ensure that the response delay error of the tactile feedback and speech synthesis and image generation modules does not exceed 50 milliseconds; When the closed-loop optimization process detects that the user generates negative emotional feedback due to interaction asynchrony, the adaptive calibration of tactile feedback parameters is automatically triggered.

6. The method for optimizing marketing strategies based on artificial intelligence dialogue according to claim 1, characterized in that: Introducing a predictive emotional state estimation mechanism, the method also includes: Based on the historical emotion trajectory of the dynamic emotion map and the current multimodal feature vector, the emotion evolution trend at the next moment is predicted through a time series neural network model to obtain the corresponding emotion state trend information; The emotional state trend information is input into the strategy generation module for adjusting the strategy generation direction in advance before actual feedback arrives, so as to enhance the emotional adaptation response to the user's potential emotional state.

7. Based on artificial intelligence dialogue marketing strategy optimization system, it is characterized by: include: The multimodal feature vector extraction module collects multi-round interaction data between the user and the AI ​​dialogue system in real time through a multimodal perception component deployed on the terminal device. The multi-round interaction data includes voice signals, text data, facial expression signals, and physiological signals. Based on this multi-round interaction data, a dynamic emotion map is generated. The nodes of the dynamic emotion map represent the user's emotional state, and the edges represent the transition probability between emotional states. The node weights are dynamically adjusted over time series to extract the multimodal feature vectors corresponding to the multi-round interaction data. The strategy risk score output module inputs the multimodal feature vector into the cross-modal information processing module, aligns the multimodal feature vector using a contrastive learning algorithm, constructs a cross-modal embedding space, and generates a causal relationship description between user behavior and marketing strategy in the cross-modal embedding space. This causal relationship description is combined with historical interaction data to simulate the possible long-term impact of different marketing methods to output the corresponding strategy risk score. A model parameter increment generation module inputs the policy risk score and causal graph model into a privacy-enhanced federated learning framework to collaboratively train the marketing strategy model across multiple enterprise nodes. Each enterprise node generates differentially private model parameter increments based on local data, encrypts the parameter increments using homomorphic encryption technology, and uploads them to the federated learning server. The federated learning server uses a secure multi-party computing protocol to aggregate model parameters based on encrypted parameter increments to generate a global policy model and updated model parameter increments. The alternative solution extraction module inputs the global strategy model and updated model parameter increments into the strategy generation module. Combined with the user's current emotional state in the dynamic emotion map, it is used to extract alternatives to high-risk strategies from the causal graph model and generate an emotionally adapted dialogue script. Simultaneously, a reinforcement learning algorithm is used to optimize strategy selection in real time to maximize the weighted objective function of user emotional satisfaction and long-term commercial value, ultimately generating personalized responses and feedback data. These personalized responses achieve multi-sensory interaction through speech synthesis, image generation, and tactile feedback devices. The update module transmits the feedback data back to the multimodal perception component for updating the dynamic emotion map and multimodal feature vectors, forming a closed-loop optimization process; the closed-loop optimization process adjusts the weights of user emotional state nodes in real time, updates the policy risk scores in the causal graph model, and dynamically adjusts the global policy model in the federated learning framework.

8. An artificial intelligence-based dialogue marketing strategy optimization system, characterized by: It includes a processor running a program of the artificial intelligence dialogue marketing strategy optimization method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: A program for the artificial intelligence dialogue marketing strategy optimization method as described in any one of claims 1 to 6 is stored.

Images